The boring take on agent memory is that coding agents forget too much.
That is true, but it is also the least interesting part of the problem.
The real issue starts when an agent stops treating context as temporary and starts turning it into durable state. A bad answer in chat is annoying. A bad memory is worse because it can quietly steer the next task, the next branch, and the next review. It becomes part of the engineering system without going through the engineering process.
That is the part people keep underestimating.
Persistent agent memory is useful. I want agents to remember project conventions, old decisions, failed approaches, deployment weirdness, naming rules, and the sharp edges that never make it into the README. Nobody wants to re-explain the same repo every morning.
